Cellular scaling laws in the mammalian brain.

Abstract

Recent whole-brain cell atlases have uncovered a consistent cytoarchitectural feature: the greatest diversity of discrete neuronal cell types resides not in the regions with the most neurons (e.g., the cortex and cerebellum) but rather in deep subcortical structures like the hypothalamus and brain stem. We propose that this discrepancy reflects a fundamental algorithmic division in the vertebrate brain between a "learning subsystem" and a "steering subsystem." The learning subsystem (cortex, striatum, and cerebellum) scales via the replication of repetitive modules to maximize computational capacity, analogous to scaling up parameters in machine learning models. By contrast, the steering subsystem (hypothalamus, pallidum, and brainstem) scales via diversification of bespoke cell types to encode innate drives and reflexes, functioning as a high-dimensional biological "reward function." This framework explains the divergence in how evolution has influenced these subsystems and offers a unified lens for understanding brain architecture and the etiology of brain disease; furthermore, it may inform model designs for artificial intelligence.